Deep Learning Models for Vascular Image Reconstruction and Analysis in Swept-Source Optical Coherence Tomography

Abstract

Optical coherence tomography (OCT) is a non-invasive imaging technique that acquires images of biological samples with micrometre resolution, thereby offering a penetration depth of a few millimetres to generate volumetric data. However, the full capabilities of OCT imaging remain underexplored in addressing challenges across many biomedical fields. Vascular studies constitute one category of these fields, where highly accurate, non-destructive imaging techniques are required to facilitate subsequent image analysis and interpretation. Thus, a thorough mechanical evaluation of the veins is crucial for understanding their functionality, designing effective venous substitutes, and gaining deeper insights into the biomechanical aspects of venous diseases. Additionally, there is considerable demand for enhancing the OCT processing pipeline to improve image quality, reduce computational complexity, and accelerate processing time. To obtain depth-resolved information, typical Swept Source OCT (SS-OCT) systems perform resampling in the wavenumber domain as a fundamental processing step. This process either demands additional hardware resources or increases the overall computational burden. With the swift evolution of deep learning (DL) techniques, automation of imaging tasks has witnessed significant progress, leading to more efficient workflows. Along this direction, a multidisciplinary approach that combines signal processing, machine learning, and wave optics is explored in this thesis to tackle key challenges and propose innovative computational methods for SS-OCT system.

In this thesis, SS-OCT was examined as an imaging modality for assessing human varicose veins, a type of venous insufficiency disease. Accurately determining the thickness of soft tissues presents a significant challenge due to their deformable and variable nature. To address this, advanced DL models are employed to capture intricate details of veins by performing segmentation and thickness estimation on SS-OCT imaging data. The core neural network architecture was an encoder-decoder model, referred to as Opto-UNet, specifically optimised to generate high-accuracy segmentation maps. Furthermore, a comprehensive model capable of generating both the segmentation maps and corresponding thickness evaluations was investigated. This model leverages transfer learning to harness the rich feature representations learned during segmentation for estimating the thickness of the region of interest in veins. A key contribution of this work is the sensitive pattern detector (SPD) module, positioned at the encoder’s output. This module enhances the network’s learning process by integrating semantic information, thereby guiding the model towards more context-aware predictions.

To tackle the high computational demands while generating high-fidelity images, this thesis analysed raw SS-OCT data, which are non-linear in the wavenumber domain and thus require resampling. The challenges within the processing pipeline were tackled using DL-based methods. Specifically, low-quality, blurred images derived from raw data were enhanced into high-quality images using a customised neural network named WAVE-UNET designed primarily for image reconstruction. Further, the image quality was enhanced by deploying two encoder-decoder styled DL models – one optimised in the spatial domain and another in the Fourier domain (FD). This dual-domain approach enables the rebuilding of impaired morphological structures in conjunction with effective noise suppression. Additionally, it significantly reduced the time complexity compared to a traditional processing pipeline.

In summary, this thesis explored SS-OCT as a promising imaging technique for evaluating soft vascular structures such as veins and analysing their mechanical characteristics. The synergy between the SS-OCT pipeline and DL demonstrates strong potential as a robust reconstruction framework. The methods developed throughout this thesis have shown high efficiency and performance, thereby advancing the state-of-the-art works in the realm of OCT.